Instruction vs. Construction: How Might Generative Artificial Intelligence Best Support Learning?

• In Event: Learning Under Algorithmic Conditions: Critical Perspectives on AI and Curriculum

Abstract

Instruction vs. Construction: How Might Generative Artificial Intelligence Best Support Learning?

Objectives
----------
Conventional learning theories typically distinguish between direct instruction and active student construction of meaning. While direct instruction emphasizes learning by showing and telling, constructionism emphasizes learning by constructing—in order to construct meaningful knowledge in one’s head, one should construct something “tangible” outside of one’s head (whether physical or digital). These two paradigms have deep roots in different approaches to AI, different approaches to studying human cognition, and different philosophies of education (Doroudi, 2023). In 2022, the world was taken aback with the release of ChatGPT, a kind of generative artificial intelligence. Generative AI is rooted in a third paradigm in the history of AI called connectionism, but this approach did not have an analogous framing in education. The question we investigate in this paper is where do generative AI tools fit in the space of curriculum studies, in terms of the dichotomy between instructionism and constructionism? By situating generative AI in the space of existing approaches to AI in education, we thereby seek to gain clarity on the role that generative AI may play in the future of education.

Theoretical Framework
---------------------
Our analysis is rooted in three sources. First, we situate different approaches to AI and education using a theoretical tool from machine learning, the bias-variance tradeoff. Doroudi (2020) previously showed how this tradeoff could be applied to debates in education research and practice. Second, we draw on the rich history of AI and education to situate different approaches (Papert, 1991). Third, we look towards technical literature on the limitations and capabilities of generative AI (West et al., 2023; Xu et al., 2024). By triangulating these three sources, we seek to understand how to situate generative AI in the space of curriculum studies.

Methods and Data
----------------
We triangulate evidence from various sources, including the bias-variance tradeoff, a conceptual discussion of limitations of generative AI, and a concrete educational case study of constructionist curriculum using generative AI, focused on mathematics teaching and learning.

Findings and Scholarly Significance
-----------------------------------
We make the case that generative AI appears to be more fundamentally better suited for constructionist applications rather than instructionist applications. In particular, we show that in terms of the bias-variance tradeoff, generative AI and its connectionist predecessors resonate more closely with constructionism. Moreover, we show that while limitations of generative AI are particularly concerning when it is being used to teach or tutor students, they are less concerning (and in some cases desirable) when it is used in constructionist ways. Finally, we describe a case study of using generative AI to help students learn Turtle geometry (Abelson and diSessa, 1986) to show the affordances of a constructionist approach. While generative AI can be used in instructionist applications, this must be done with care and might require creative ways of integrating generative AI with symbolic AI. But we should also consider the role that generative AI can play in the future of education as powerful constructionist microworlds.

Authors

• Shayan Doroudi, University of California - Irvine

• Sina Rismanchian, University of California - Irvine